More than 13,000 colleagues are participating in Deutsche Telekom's World Cup betting pool. And a very special better is participating at Telekom Innovation Laboratories (T-Labs): an AI (artificial intelligence) bot programmed in-house that is trying to pick the winners of all the World Cup matches. Without any emotion, based solely on data from previous World Cup tournaments since 1930. But the exit of the German team from the tournament has predicted the AI as little as most of us.
Quite the contrary: The AI has picked Germany to reach the final. But with this prognosis, the AI system was certainly not alone. But let’s have a look on the T-Labs oracle besides the German result: Who can beat the AI in the betting pool? "We're really proud of our AI. After the group phase, it's close to the top of the rankings," explain Ronald Fromm and Elmar Arunov, who both work in the artificial intelligence innovation area at T-Labs. In the T-Labs pool group, the AI is currently only three points behind the top 3. Not bad for a group with 60 participants in total. World Cup newcomer Panama proved to be a setback for the AI, because this is the first time the team is participating. That means no World Cup data is available for Panama, which the AI needs for its predictions. As such, the AI predicted the Panama-England match would end 1:1, instead of the actual 6:1 result for England – a wrong prediction, as it turned out. In contrast, the AI predicted Germany's last minute 2:1 victory over Sweden on the nose. It also predicted several ties correctly, including Spain versus Morocco, although results without a victor are the most difficult to prognosticate.
T-Labs established artificial intelligence as an innovation area around a year ago. The team focuses mainly on the use of AI methods, such as machine learning, in network-relevant use cases. "The idea to use the AI for the World Cup arose from a student project," says Elmar Arunov. "Maik Pächter, one of our cooperative students, designed and trained the model with our support." All the World Cup data since 1930 was "fed in", which only took a couple of hours. The AI did not get any additional information, however, such as results from national leagues or current injuries. The goal of the project was simply to see the best way to train AI-based models optimally with a minimum of input data.
The bottom line? "The AI is predicting the results as good and as reliably as a regular soccer expert. It will be interesting to see where it ends up in the end," says Elmar Arunov. "But it has great faith in the Brazilian team." If the T-Labs oracle is to be believed, Brazil will win this year’s final match of the World Cup. Incidentally, around 90 percent of other neural networks that have made predictions also picked Brazil to win the title.
T-Labs is now constructing a new model for the knockout matches. This will enable them to determine whether the AI has learned from the results of the group round, possibly changing its prediction. In that regard, watching the AI develop might be just as exciting as the play on the field.